Weight sharing for machine learning models on heterogeneous cores

Abstract

Systems and techniques are described herein for running a machine learning (ML) model. For instance, a process can include loading a data structure from a long-term memory into a short-term memory, the data structure including embedding information for mapping identifiers to corresponding vectors; transposing the embedding information to generate a transposed data structure; and providing the transposed data structure to a language model (LM) head for translating feature vectors to the identifiers.

Images